Skip to content

Mitigating Biases and Prompt Effects in Vision-Language Models #75

@gautamjajoo

Description

@gautamjajoo

Project Title: Mitigating Biases and Prompt Effects in Vision-Language Models

Description: This research project aims to investigate how prompt engineering influences the behaviour and outputs of Vision-Language Models (VLLMs), with a particular focus on the emergence and amplification of biases. By systematically studying the relationship between prompt formulations and model responses, the project seeks to uncover the mechanisms through which biases are introduced and propose effective mitigation strategies.

Deliverables:

  • Literature Review: Conduct an extensive review of current research on prompt engineering, bias in AI, and VLLM behavior.
  • Experimental Design: Set up experiments where prompt variables are systematically manipulated, and corresponding outputs are analyzed.
  • Bias Taxonomy: Create a taxonomy of biases that emerge from prompt variations in VLLMs, including demographic, cultural, and contextual biases.
  • Cross-Model Comparison: Analyze and compare the sensitivity of different VLLM architectures to prompt-induced biases to identify common patterns and model-specific vulnerabilities.
  • Benchmarking: Compare the performance of baseline models with those employing mitigation strategies across diverse scenarios and prompt configurations.

Mentors: @gautamjajoo @RishabhJain2018

Skills: Python, PyTorch/TensorFlow, NLP, Vision-Language Models, LLMs, Bias and Fairness in AI

Skill Level: Medium to Advanced

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions